Searching for a Middle Term

“But nothing, I think, prevents one from in a sense understanding and in a sense being ignorant of what one is learning” (Aristotle, Posterior Analytics; Complete Works revised Oxford edition vol. 1, p. 115). The kind of understanding spoken of here involves awareness “both that the explanation because of which the object is is its explanation, and that it is not possible for this to be otherwise” (ibid). To speak of the “explanation because of which” something is suggests that the concern is with states of affairs being some way, and the “not… otherwise” language further confirms this.

Following this is the famous criterion that demonstrative understanding depends on “things that are true and primitive and immediate and more familiar than and prior to and explanatory of the conclusion…. [T]here will be deduction even without these conditions, but there will not be demonstration, for it will not produce understanding” (ibid). The “more familiar than” part has sometimes been mistranslated as “better known than”, confusing what Aristotle carefully distinguishes as gnosis (personal acquaintance) and episteme (knowledge in a strong sense). I think this phrase is the key to the whole larger clause, giving it a pragmatic rather than foundationalist meaning. (Foundationalist claims only emerged later, with the Stoics and Descartes.) The pedagogical aim of demonstration is to use things that are more familiar to us — which for practical purposes we take to be true and primitive and immediate and prior and explanatory — to showcase reasons for things that are slightly less obvious.

Independent of these criteria for demonstration, the whole point of the syllogistic form is that the conclusion very “obviously” and necessarily follows, by a simple operation of composition on the premises (A => B and B => C, so A=> C). Once we have accepted both premises of a syllogism, the conclusion is already implicit, and that in an especially clear way. We will not reach any novel or unexpected conclusions by syllogism. It is a kind of canonical minimal inferential step, intended not to be profound but to be as simple and clear as possible.

(Contemporary category theory grounds all of mathematics on the notion of composable abstract dependencies, expressing complex dependencies as compositions of simpler ones. Its power depends on the fact that under a few carefully specified conditions expressing the properties of good composition, the composition of higher-order functions with internal conditional logic — and other even more general constructions — works in exactly the same way as composition of simple predications like “A is B“.)

Since a syllogism is designed to be a minimal inferential step, there is never a question of “searching” for the right conclusion. Rather, Aristotle speaks of searching for a “middle term” before an appropriate pair of premises is identified for syllogistic use. A middle term like B in the example above is the key ingredient in a syllogism, appearing both in the syntactically dependent position in one premise, and in the syntactically depended-upon position in the other premise, thus allowing the two to be composed together. This is a very simple example of mediation. Existence of a middle term B is what makes composition of the premises possible, and is therefore what makes pairings of premises appropriate for syllogistic use.

In many contexts, searching for a middle term can be understood as inventing an appropriate intermediate abstraction from available materials. If an existing abstraction is too broad to fit the case, we can add specifications until it does, and then optionally give the result a new name. All Aristotelian terms essentially are implied specifications; the names are just for convenience. Aristotle sometimes uses pure specifications as “nameless terms”.

Named abstractions function as shorthand for the potential inferences that they embody, enabling simple common-sense reasoning in ordinary language. We can become more clear about our thinking by using dialectic to unpack the implications of the abstractions embodied in our use of words. (See also Free Play; Practical Judgment.)

Aristotelian Propositions

Every canonical Aristotelian proposition can be interpreted as expressing a judgment of material consequence or material incompatibility. This may seem surprising. First, a bit of background…

At the beginning of On Interpretation, Aristotle says that “falsity and truth have to do with combination and separation” (Ch. 1). On its face, the combination or separation at issue has to do not with propositions but with terms. But it is not quite so simple. The terms in question are canonically “universals” or types or higher-order terms, each of which is therefore convertible with a mentioned proposition that the higher-order term is or is not instantiated or does or does not apply. (We can read, e.g., “human” as the mentioned proposition “x human”.) Thus a canonical Aristotelian proposition is formed by “combining” or “separating” a pair of things that are each interpretable as an implicit proposition in the modern sense.

Propositions in the modern sense are treated as atomic. They are often associated with merely stipulated truth values, and in any case it makes no sense to ask for internal criteria that would help validate or invalidate a modern proposition. But we can always ask whether the combination or separation in a canonical Aristotelian proposition is reasonable for the arguments to which it is applied. Therefore, unlike a proposition in the modern sense, an Aristotelian proposition always implicitly carries with it a suggestion of criteria for its validation.

The only available criteria for critically assessing correctness of such elementary proposition-forming combination or separation are material in the sense that Sellars and Brandom have discussed. A judgment of “combination” in effect just is a judgment of material consequence; a judgment of “separation” in effect just is a judgment of material incompatibility. (This also helps clarify why it is essential to mention both combination and separation affirmatively, since, e.g., “human combines with mortal” canonically means not just that human and mortal are not incompatible, but that if one is said to be human, one is thereby also said to be mortal.)

This means that Aristotle’s concept of the elementary truth and falsity of propositions can be understood as grounded in criteria for goodness of material inference, not some kind of correspondence with naively conceived facts. It also means that every Aristotelian proposition can be understood as expressing a judgment of material consequence or incompatibility, and that truth for Aristotle can therefore be understood as primarily said of good judgments of material consequence or incompatibility. Aristotle thus would seem to anticipate Brandom on truth.

This is the deeper meaning of Aristotle’s statement that a proposition in his sense does not just “say something” but “says something about something”. Such aboutness is not just grammatical, but material-inferential. This is in accordance with Aristotle’s logical uses of “said of”, which would be well explained by giving that a material-inferential interpretation as well.

The principle behind Aristotelian syllogism is a form of composition, formally interpretable as an instance of the composition of mathematical functions, where composition operates on the combination or separation of pairs of terms in each proposition. Aristotelian logic thus combines a kind of material inference in proposition formation and its validation with a kind of formal inference by composition. This is what Kant and Hegel meant by “logic”, apart from their own innovations.

Propositions, Terms

Brandom puts significant emphasis on Kant and Frege’s focus on whole judgments — contrasted with simple first-order terms, corresponding to natural-language words or subsentential phrases — as the appropriate units of logical analysis. The important part of this is that a judgment is the minimal unit that can be given inferential meaning.

All this looks quite different from a higher-order perspective. Mid-20th century logical orthodoxy was severely biased toward first-order logic, due to foundationalist worries about completeness. In a first-order context, logical terms are expected to correspond to subsentential elements that cannot be given inferential meaning by themselves. But in a higher-order context, this is not the case. One of the most important ideas in contemporary computer science is the correspondence between propositions and types. Generalized terms are interpretable as types, and thus also as propositions. This means that (higher-order) terms can represent instances of arbitrarily complex propositions. Higher-order terms can be thus be given inferential meaning, just like sentential variables. This is all in a formal context rather than a natural-language one, but so was Frege’s work; and for what it’s worth, some linguists have also been using typed lambda calculus in the analysis of natural language semantics.

Suitably typed terms compose, just like functions or category-theoretic morphisms and functors. I understand the syllogistic principle on which Aristotle based a kind of simultaneously formal and material term inference (see Aristotelian Propositions) to be just a form of composition of things that can be thought of as functions or typed terms. Proof theory, category theory, and many other technical developments explicitly work with composition as a basic form of abstract inference. Aristotle developed the original compositional logic, and it was not Aristotle but mid-20th century logical orthodoxy that insisted on the centrality of the first-order case. Higher-order, compositionally oriented logics can interpret classic syllogistic inference, first-order logic, and much else, while supporting more inferentially oriented semantics on the formal side, with types potentially taking pieces of developed material-inferential content into the formal context. We can also use natural-language words to refer to higher-order terms and their inferential significance, just as we can capture a whole complex argument in an appropriately framed definition. Accordingly, there should be no stigma associated with reasoning about terms, or even just about words.

In computer-assisted theorem-proving, there is an important distinction between results that can be proved directly by something like algebraic substitution for individual variables, and those that require a more global rewriting of the context in terms of some previously proven equivalence(s). At a high enough level of simultaneous abstraction and detail, such rewriting could perhaps constructively model the revision of commitments and concepts from one well-defined context to another.

The potential issue would be that global rewriting still works in a higher-order context that is expected to itself be statically consistent, whereas revision of commitments and concepts taken simply implies a change of higher-level context. I think this just means a careful distinction of levels would be needed. After all, any new, revised genealogical recollection of our best thoughts will be in principle representable as a new static higher-order structure, and that structure will include something that can be read as an explanation of the transition. It may itself be subject to future revision, but in the static context that does not matter.

The limitation of such an approach is that it requires all the details of the transition to be set up statically, which can be a lot of work, and it would also be far more brittle than Brandom’s informal material inference. (See also Categorical “Evil”; Definition.)

I am fascinated by the fact that typed terms can begin to capture material as well as purely formal significance. How complete or adequate this is would depend on the implementation.